The rapid growth of the Internet of Things (IoT) and the ubiquitous nature of the Internet have made life more convenient for human beings. The rise of that social convenience is accompanied by incessant efforts of miscreants to create new tools, techniques, and tactics to destabilize the comfort of the dwellers by attacking computer networks and applications. Even worse, these attacks are being transferred into the increasingly connected cyber-physical systems (CPS), especially the automotive system where the in-vehicle CAN bus network lacks encryption and authentication mechanisms, making them even more vulnerable to some of the attacks that are well-known in traditional computer networks. Additionally, some automotive systems (e.g., the modern car) employ advanced technologies ?the Telematics Unit, in-vehicle infotainment (IVI), V2X, etc.?accessible through Bluetooth, Wi-Fi, GPS, etc., thus, augmenting their attack surface. Intrusion Detection Systems are known to be the solution by excellence for detecting and mitigating network attacks, however, based on the recrudescence of attacks, we can affirm that traditional IDSs have failed. Elsewhere, artificial intelligence (AI) or, more specifically, deep learning has shown immense promise in solving lingering issues in other domains: we contend that deep learning can also help make IDSs more efficient.
Hence, in this dissertation, our imperative is to devise new IDS methodologies to protect computer networks and in-vehicle CAN bus networks of automotive systems by leveraging deep learning. First, we thoroughly study the deep learning-based IDS for several kinds of critical network attacks such as DoS (Denial of Service), DDoS (Distributed DoS), Brute Force, etc. Subsequently, we investigate how to optimize the deep learning models. Our results illustrate that Long Short-Term Memory (LSTM) can effectively detect network attacks with high accuracy and reasonable detection rates. After ensuring security in computer networks by using deep learning, like the attackers, we also transfer our solutions to the automotive systems. Therefore, we propose a deep learning-based IDS for in-vehicle CAN bus networks. Furthermore, for efficiency reasons, we also develop CAN bus network attacks (DoS, Fuzzing, and Spoofing) datasets by using the CAN messages of three distinct car models (Toyota, Subaru, and Suzuki). The results of our experiment demonstrate that our deep learning-based IDS is more effective and robust than existing methodologies.